Re: FreeBSD 8.0-BETA2/amd64 crashes on SMP under load

On Friday 31 July 2009 2:01:49 pm Marcel Moolenaar wrote:
> 
> On Jul 31, 2009, at 10:25 AM, John Baldwin wrote:
> >> In other words: kgdb was designed to hide FreeBSD specifics from the
> >> core GDB
> >> code, because core GDB doesn't know how to deal with all FreeBSD
> >> details.
> >> Revision 178670 created a stronger dependency on core GDB and as such
> >> broke
> >> architectures that core GDB doesn't support for FreeBSD.
> >
> > So gdb doesn't work for core dumps on regular processes under ia64?
> 
> It does, but kernel core files aren't matched, because the OSABI
> is standalone. Also, GDB doesn't have libkvm support for ia64,
> which is probably what makes it work on i386/amd64.

Regular gdb doesn't support libkvm on i386/amd64 either.

> >  Still,
> > kgdb doesn't actually use the regular core target at all, it uses  
> > its own
> > target that uses libkvm to service the actual memory I/O that  
> > normally goes
> > to the core target.
> 
> That's maybe the problom: is the kgdb target used at all?
> If it were, then why did ia64 break?
> 
> Maybe we have a simple resolution problem where GDB just
> happens to pick the wrong core_stratum target on ia64...

Hmm, so kgdb actually forces its own core_stratum target always via 
initalize_kgdb_target().  It overrides the "base" core_target that uses 
libbfd, etc.  Thus, kgdb will always use libkvm and not libbfd for handling 
core files.

> I haven't had the time for root cause hunting, so from here
> on things should be taken with plenty of salt. I can check
> PowerPC for example...

I do believe it has been used on sparc64 at least since my changes.

> >  The previous code was very hacky and didn't allow you
> > to use the 'file' and 'core' commands as a result.
> 
> Actually it was by design.

Well, that wasn't very intuitive to folks used to regular gdb.  I think it now 
functions more like gdb in that you can switch executables or core images and 
modules show up as shared libraries, etc.  IIRC, it was also a lot harder to 
add KLD/module support in the previous arrangement since I needed to override 
the thread_stratum for kld handling independent of having either a remote 
target or a vmcore target.  That is probably the main reason I changed it 
from thread_stratum to core_stratum actually.

-- 
John Baldwin